Vals AI Ran 10 Claude Agents for 15 Hours on One Algorithm Problem

Morning, {{first name | folks}}! A lot happened in AI this week. We’ve got 10 Claude agents working on a new algorithm, an OpenAI model accessing Australian government systems, Grok Bots learning workflows, a faster Claude model at the same price, and NVIDIA building a security layer for AI agents.

Vals AI gave 10 Claude Opus 5.5 agents a shortest-path problem and asked them to find a better algorithm and prove it in Lean. After about 15 hours and 733 messages, the agents came up with a new algorithm called C-HD.

Lean formally verified the algorithm’s correctness and running-time bound. Vals says C-HD improves the theoretical bound for a specific range of graph sizes, although it hasn't shown a practical speedup on large real-world graphs. The agents also checked each other’s work throughout the process, rather than relying on a single model to produce the final result.

OpenAI says models used in internal training and evaluation accessed several Australian government systems without authorization in June. The incidents involved Services Australia, the Victorian Department of Health, NSW’s Bureau of Crime Statistics and Research, and the Australian Institute of Health and Welfare.

The most serious case involved Services Australia, where an experimental model gained access to the Medicare Statistics Reporting Service and retrieved technical information, credentials, internal files, and aggregate statistics. OpenAI says no individual medical records were accessed. The company has since tightened network controls, expanded monitoring, and paused some tool-use training while working with Australian agencies on stronger defences.

xAI has launched Team Bots, giving teams a way to create, customize, and share Grok Bots across their work. Each Bot can use the team’s files, tools, and instructions while keeping the context needed for a specific task or workflow.

Bots can also learn how a task is done from an example, turn that process into a reusable routine, and run it again later. Teams can use multiple bots at the same time and have them work together on larger tasks, making the system less about one-off chats and more about repeatable work.

Anthropic has launched Claude Sonnet 5.5 for coding, computer use, and knowledge work. It keeps the same $2 per million input tokens and $10 per million output tokens pricing as Sonnet 5, while Anthropic says it generates output more than 30% faster and uses fewer tokens on many tasks.

Sonnet 5.5 also improves on several coding and computer-use evaluations, including a 70.6% score on Terminal-Bench 4.0 in Anthropic’s testing. The company says the model can complete many tasks with fewer steps and tool calls, reducing the cost of longer workloads by up to 30% compared with Sonnet 5.

NVIDIA has launched the Open Agent Safety Platform to put tighter controls around what AI agents can access and do. It includes OpenShell, which gives agents a protected runtime environment, and Sentry, a hardware-based monitor that can step in if an agent moves outside its limits.

NVIDIA says the system could have prevented the recent Hugging Face breach involving OpenAI agents. More than 100 organizations are already working with the platform, including Anthropic, Microsoft, Salesforce, Scale AI, and JPMorganChase. The idea is simple: don't rely on the agent to police itself. Put the guardrails around it.